Install
Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/dgenio/contextweaver/agents-mdgit clone --depth 1 https://github.com/dgenio/contextweaver Per session 17,577 This file is loaded in full into every session.
When invoked 17,577 The same file — it is already loaded in full.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.17577 | $0.17577 |
| Opus 5 | $0.08788 | $0.08788 |
| Sonnet 5 | $0.03515 | $0.03515 |
| Haiku 4.5 | $0.01758 | $0.01758 |
Security
Grade A, and why
contextweaver AGENTS.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
AGENTS.md · 413 lines
How it starts
The opening of the file, as written. The whole thing — 413 lines — stays where its author put it; the contents beside it link to each section on GitHub.
contextweaver — Agent Guide
Authority: This file is the single source of truth for agent-facing guidance. Tool-specific files (
.claude/CLAUDE.md,.github/copilot-instructions.md) contain only overrides and guardrails — they defer to this file for all shared rules.
Purpose
contextweaver is a Python library for dynamic context management for tool-using AI agents. It provides two integrated engines:
- Context Engine — phase-specific budgeted context compilation with a context firewall
- Routing Engine — bounded-choice navigation over large tool catalogs via DAG + beam search
Non-goals: contextweaver is not an LLM inference layer and not a tool execution runtime. It prepares context and routes tools but never calls models or executes tools.
Module Map
| Path | Responsibility |
|---|---|
types.py |
Core dataclasses and enums: SelectableItem, ContextItem, Phase, ItemKind, Sensitivity |
envelope.py |
Result types: ResultEnvelope, BuildStats, DroppedItem, ContextPack, ChoiceCard, HydrationResult, RoutingDecision |
diagnostics.py |
Versioned, payload-safe gateway events and sinks (DiagnosticEvent, DiagnosticSink, JSONL/in-memory sinks) plus deterministic aggregate reports (issues #370/#378). |
inspection.py |
Pure JSON/Markdown report construction for offline context, routing, and artifact inspection without raw payload content (issue #398). |
telemetry_contract.py |
Versioned JSONL telemetry handoff contract over DiagnosticEvent for downstream analytics (issue #382): EVENT_FAMILIES (8 families → event-name prefixes; shortlist/policy_denial/visibility reserved until their emitters land), classify_event, validate_event_dict (envelope shape + payload-leak heuristic), export_jsonl/read_jsonl (skip-and-collect). Hand-written envelope schema under schemas/telemetry/v1/; contract doc docs/telemetry.md. |
ops_view.py |
Read-only gateway ops-triage view over the diagnostics JSONL (issue #668): windowed OpsSnapshot (family counts, error rate, browse/execute latency percentiles, top executed/failing tools), plain-text + Rich renders, incremental file-tailing watch_loop. Backs mcp ops. |
visualize.py |
Self-contained HTML reports for RouteTrace / BuildStats / diagnostic-event timelines (issue #442): inline CSS only, all dynamic (untrusted) text HTML-escaped, byte-identical output for identical inputs. Backs the visualize CLI command. |
config.py |
Configuration: ContextBudget, ContextPolicy (incl. overflow_action budget-overflow policy, issue #510), and ScoringConfig |
profiles.py |
Routing and profile config: Mode, RoutingConfig, ProfileConfig, named presets |
protocols.py |
Protocol interfaces: TokenEstimator, EventHook, Summarizer, Extractor, RedactionHook, SensitivityClassifier (ingestion-time labelling, issue #542), MemorySource, Labeler, Retriever, Reranker, ClusteringEngine, RoutingScoreProvider (store protocols re-exported from store/protocols.py). Bundled estimators: HeuristicEstimator (default, script-aware, dependency-free — counts CJK/Kana/Hangul/emoji ≈1 token/char, issue #525), CharDivFourEstimator (raw len // 4 primitive), TiktokenEstimator (exact, falls back to HeuristicEstimator offline). Each carries a stable name for BuildStats.token_estimator. |
store/protocols.py |
Store-layer protocols: EventLog, ArtifactStore, EpisodicStore, FactStore |
store/async_protocols.py |
Async counterparts AsyncEventLog / AsyncArtifactStore / AsyncEpisodicStore / AsyncFactStore (issue #495) — same surface, async def. Consumed only by the async context/ path; backend-agnostic. |
store/async_bridge.py |
to_async(sync_store) — wraps a thread-safe sync backend as the matching async protocol via asyncio.to_thread. Thread-affine backends (SqliteEventLog, check_same_thread=True) are not valid targets (issue #495). |
store/_async_to_sync.py |
Inverse bridges + to_sync(async_store, loop) + is_async_store() (issue #495). Drives async stores on a private _LoopThread so the existing sync pipeline can consume them; ContextManager offloads build to a worker thread when async-backed. Not public API. |
store/_loop_thread.py |
The private _LoopThread (asyncio loop on a daemon thread) backing _async_to_sync (kept separate to isolate loop lifecycle and timeout handling). run(coro, timeout=…) bounds each op and raises StoreTimeoutError instead of hanging a stuck backend (issue #750). Not public API. |
exceptions.py |
Custom exception hierarchy (all errors inherit ContextWeaverError). Each class carries a stable, frozen code (e.g. CW_CONFIG) plus an optional hint; str(exc) renders [code] message (hint: …). Codes are documented in docs/errors.md and golden-listed in tests/test_exceptions.py (issues #635, #637). |
_utils.py |
Text similarity primitives: tokenize(), jaccard(), TfIdfScorer |
secrets.py |
Pure, deterministic secret detection/scrubbing primitives: scrub_secrets(), scrub_secrets_in_list(), contains_secret(), SecretPattern (issue #428). Shared by the firewall secret-scrub, the SecretRedactor hook, the sensitivity classifier, and ChoiceCard scrubbing. No I/O; never weakens a surface (only removes characters). |
_version.py |
Single-source version derived from importlib.metadata; fallback "0.0.0+local" |
_incident_pack.py |
Offline incident-pack builder for contextweaver mcp incident-pack (issue #661): creates bounded redacted zip bundles with a machine-readable manifest, summaries, redacted config/catalog/diagnostics excerpts, and reproduction checklist. Private; no public API. |
_incident_pack_files.py |
Private archive-entry helpers for incident packs: key-aware structured redaction, per-file truncation, hashes/timestamps for manifest entries, and best-effort JSON/YAML/JSONL parsing. Not public API. |
_vscode_import.py |
Pure transform for contextweaver mcp import-vscode (issue #367): build_migration_plan converts a VS Code-family MCP config's servers/mcpServers block into an upstreams: gateway config; render_gateway_config/render_replacement_config render the two output configs (kept separate from the plan since the replacement config embeds a write-time --gateway-config path decision); write_migration applies with a default backup. Not public API. |
_demos.py |
Demo logic for the CLI demo subcommand (exempt from print() rule) |
serde.py |
Serialisation helpers for to_dict / from_dict |
tokens.py |
Built-in token counter (count(), get_token_counter(), heuristic_counter(), TokenCounter alias) plus the provider-estimator registry (register_estimator(), registered_estimators(), estimator_name(), issue #493). The single source of truth for token counts — firewall, sensitivity-redaction placeholders, card budgeting (routing/cards.count_tokens), and FirewallStats/BuildStats numbers all route through it (issues #405/#493/#530); no stray len // 4 literals elsewhere. Owns the tiktoken dependency; offline it falls back to the script-aware HeuristicEstimator. |
store/ |
In-memory data stores: EventLog, ArtifactStore, EpisodicStore, FactStore, StoreBundle |
store/_sqlite_base.py |
Shared SQLite connection + migration scaffolding (WAL, foreign_keys=ON, _contextweaver_schema_version table). Reused by every SQLite-backed store (issue #174). |
store/sqlite_event_log.py |
SqliteEventLog — first persistent EventLog backend; single-process, sync, append-only, schema-versioned (issue #223). |
store/sqlite_episodic.py |
SqliteEpisodicStore — persistent EpisodicStore on _sqlite_base (issue #496). Append-only, ordered by ordinal; search delegates to a transient InMemoryEpisodicStore for byte-identical ranking. Own version table (VERSION_TABLE) so it can share a DB file with the event log / facts. |
store/sqlite_facts.py |
SqliteFactStore — persistent FactStore on _sqlite_base (issue #496). put upserts on fact_id; get_by_key/all sorted by fact_id. Own version table; shareable DB file. |
store/redis_artifacts.py |
RedisArtifactStore — Redis ArtifactStore for multi-process gateways (issue #426). Namespaced keys, optional per-artifact TTL, list_refs via SCAN. Lazy redis import ([redis] extra). |
store/redis_event_log.py |
RedisEventLog — Redis EventLog (issue #426). Items in a hash keyed by id + parallel order list; append-only ordering across processes. Lazy redis import ([redis] extra). |
store/s3_artifacts.py |
S3ArtifactStore — S3-compatible ArtifactStore (issue #426; AWS/MinIO/R2/GCS). {prefix}/{handle}.data + .json objects. Lazy boto3 import ([s3] extra). |
store/json_file_artifacts.py |
JsonFileArtifactStore — filesystem ArtifactStore backend; {enc(handle)}.data + {enc(handle)}.json per artifact, re-instantiable against an existing directory (issue #42). Hardened (issue #497): atomic writes (temp file + os.replace), an in-memory handle→ref index built once on init so list_refs never rescans the directory, and optional max_bytes / max_artifacts quotas raising ArtifactStoreQuotaError. Persists content_hash and percent-encodes handles into filenames so the firewall's artifact:result:… handles are Windows-safe (issue #466). Optional ttl_seconds / redact_secrets lifecycle policy (issue #375) delegates to store/_json_file_ttl.py's ArtifactLifecycle. |
store/_json_file_ttl.py |
Private TTL + redaction-before-store bookkeeping for json_file_artifacts.py (issue #375): ArtifactLifecycle (composed, not inherited) — prepare() scrubs UTF-8 content via contextweaver.secrets.scrub_secrets before write, record_put()/is_expired()/forget() manage a process-lifetime-scoped expiry dict under an injectable monotonic Clock (mirrors gateway_controls.ToolResultCache's clock convention). Not persisted across a restart. Not public API. |
store/_json_file_io.py |
Private filesystem helpers for json_file_artifacts.py (keeps filesystem concerns isolated from store policy): on-disk suffix constants, validate_handle (path-traversal defense), encode_handle (percent-encoding), and the atomic_write primitive (issues #466/#497). Not public API. |
store/testing.py |
Store-protocol conformance kit (issue #520): framework-agnostic check_event_log_conformance / check_artifact_store_conformance / check_episodic_store_conformance / check_fact_store_conformance, each taking a factory for an empty backend and asserting the round-trip / ordering / not-found contract. No test-framework import; ships in the core wheel. tests/test_store_conformance.py runs every bundled backend through it. |
summarize/ |
SummarizationRule, RuleEngine, extract_facts() |
summarize/structured.py |
Lossless JSON field projection for the firewall: parse_path / project + StructuredFirewall(keep=[...]). Deterministic, no LLM — keeps an allow-list of JSON paths inline and offloads the rest (issue #406). |
context/ |
Full context pipeline, sensitivity enforcement, view registry, ContextManager |
context/firewall_api.py |
Single-call firewall facade: compact_tool_result / firewalled_tool_result → CompactResult. Composes structured/text strategies, schema-preserving pass-through (reserved _cw sidecar — a caller payload already using _cw raises ConfigError unless overwrite_sidecar=True, #467), the built-in token counter, and fail-closed deterministic mode (issues #399, #402, #403, #404, #405, #406, #467). |
context/manager.py |
ContextManager — thin orchestrator (__init__, properties, drilldown, mixin composition). Public method stubs live in flat partial-class mixins; pipeline logic lives in the delegate modules below (issue #101). |
context/_manager_base.py |
_ManagerState — private-attribute + _build contract the manager mixins inherit and the delegate pipeline modules type their manager parameter against (ContextManager inherits it via the mixins). Not public API (issue #101). |
context/_manager_ingest.py / _manager_build.py / _manager_routing.py |
Private partial-class mixins that preserve ContextManager's public ingestion, build, and route/call-prompt API while delegating implementation to focused modules. Their composition is private implementation detail; the public behavior/API is the contract. |
context/ingest.py |
Tool-result ingestion helpers kept separate from the public manager façade to isolate ingestion concerns. Includes ingest_envelope — the canonical Frame-shaped seam (weaver-spec I-05) that ingests an already-firewalled ResultEnvelope without re-deriving firewalling; raw-output ingest_tool_result / ingest_mcp_result are non-canonical for spec compliance (issue #352). |
context/memory_types.py |
MemoryEntry dataclass + PHASE_SCOPE_PREFERENCES constants for phase-aware memory ingestion (issue #293). |
context/memory_fixture.py |
JsonFixtureMemorySource — deterministic stdlib fixture adapter implementing the MemorySource Protocol from protocols.py (issue #293). |
context/memory_source.py |
memory_entries_to_context_items / select_memory_for_phase helpers that materialise memory entries into budgeted memory_fact candidates (issue #293). |
context/handoff_types.py |
HandoffEntry + SessionHandoffPack dataclasses and canonical handoff category constants (issue #294). |
context/handoff.py |
build_session_handoff_pack / render_handoff_pack — deterministic, budget-aware, sensitivity- and firewall-respecting session continuity snapshot (issue #294). |
context/advisor_pack.py |
Advice-only planning escalation (issue #741): AdvisorRequest / AdvisorResponse + ask_advisor package a bounded, budget-truncated, strict-JSON prompt to a stronger model via call_fn (optionally GuardedCallFn-wrapped). Off-list preferred options are nulled with a deterministic marker; malformed output degrades to raw-text advice. Advice never executes or authorizes anything. |
context/consolidation_types.py |
ConsolidationPolicy / EpisodeCluster / PromotedFact / ConsolidationReport (+ CONSOLIDATION_REPORT_VERSION) — pure-data config and result types for the memory consolidation engine (issue #498). |
context/consolidation.py |
Memory consolidation engine (issue #498): cluster_episodes (deterministic episodic clustering/dedupe, #679), promote_clusters (fact promotion with provenance + max-sensitivity inheritance, #680; optional fail-closed call_fn merge, #682), decay_episodes / decay_facts (report-only decay over append-only stores, #681), and the consolidate(...) orchestrator → ConsolidationReport. Deterministic; apply=True upserts content-addressed facts (idempotent). Standalone functions (not a ContextManager method) mirroring handoff.py. |
context/_consolidation_merge.py |
Private optional model-assisted canonicalizer for consolidation (issue #682): refine_canonical_text runs a user-supplied call_fn under fail-closed guardrails (no LLM SDK dep; rejects blank/ungrounded completions that introduce tokens absent from the source cluster, falling back to the deterministic text). Not public API. |
context/explanation.py |
ContextBuildExplanation + CandidateExplanation opt-in debug surface returned by ContextManager.build(..., explain=True) (issue #291); carries resolved_weights (the per-phase scoring weights applied, issue #487). Sister to routing/explanation.py on the routing side. |
context/build_policy.py |
Pure build-pipeline policy helpers (not public API): override_phase_budget / adjust_budget_for_header (budget math), enforce_overflow_policy (ContextPolicy.overflow_action, issue #510), and render_pack_prompt (caller-owned renderer hook, issue #410). Extracted from build.py to keep it within its size ceiling. |
context/classify.py |
Opt-in deterministic ingestion-time sensitivity classification (issue #542): HeuristicSensitivityClassifier (implements the SensitivityClassifier protocol) + detect_sensitivity(). Runs at the start of the pipeline's sensitivity stage and over fact/episode header content; may only raise a label, never lower it. Reuses secrets.contains_secret plus PII markers. |
context/secret_redaction.py |
Opt-in SecretRedactor RedactionHook (issue #428): substring-scrubs secret shapes from an item's text via secrets.scrub_secrets. Registered under the name "secret" for ContextPolicy.redaction_hooks; complements (does not replace) MaskRedactionHook. |
routing/ |
Catalog, ChoiceGraph, TreeBuilder, Router (beam search), card renderer |
routing/filters.py |
Pre-scoring helpers: filter_items(), augment_query(), suggest_clarifying_question() (issues #14, #22, #112, #116) |
routing/catalog_metadata.py |
Catalog inventory/governance metadata (issue #377): InventoryMetadata (owner/domain/risk/side-effects/lifecycle/environment/tier, all default None = unknown) under the reserved metadata["_contextweaver"]["inventory"] namespace via attach_inventory / inventory_of / validate_inventory / is_deprecated. Round-trips through SelectableItem serde untouched. |
routing/collision.py |
Collision & duplicate-capability analyzer (issue #381): analyze_collisions → CollisionReport (exact/near-name, description- and schema-similar findings + recommendation heuristic + render_markdown). Similarity via _utils + stdlib difflib; small schemas skipped for false-positive suppression. |
routing/catalog_diff.py |
Catalog diff + routing impact (issue #514): diff_catalogs (added/removed/field-level changes anchored by compute_catalog_hash), routing_impact (top-1 flips + recall@k over probe queries), suggest_probes heuristic probe generator. Backs catalog diff. |
routing/vector_index.py |
Embedding index over tool catalogs (issue #387): canonical_tool_text + VectorIndex (build / hash-based incremental refresh / query with per-section evidence / cosine duplicates). Backend-agnostic over EmbeddingBackend; deterministic with HashingEmbeddingBackend; plain-list vectors. |
routing/export.py |
Provider-native tool exporters for routed shortlists (issue #609): to_openai_tools / to_anthropic_tools / to_gemini_function_declarations → ExportedTools (wire-ready tools + deterministic sanitised-name→tool_id map with collision handling). Hydrates empty inline schemas via Catalog.hydrate (inline wins). |
routing/manifest.py |
GraphManifest + compute_catalog_hash() for graph metadata and cache invalidation (issue #48, #15) |
routing/normalizer.py |
CatalogNormalizer + NormalizationReport for catalog metadata hygiene (issue #44) |
routing/catalog.py (validation) |
validate_references / Catalog.validate_references → CatalogValidationReport of dangling depends_on/requires refs; loaders take on_invalid ("warn"/"raise"/"ignore") and raise CatalogValidationError in raise mode (issue #519) |
routing/registry.py |
EngineRegistry and bundled TfIdfRetriever / NoOpReranker / JaccardClusteringEngine defaults (issue #47) |
routing/index_cache.py |
Persistent, reusable fitted-index cache (issues #543/#624/#685): RoutingIndexCache (in-process LRU + optional deterministic-JSON on-disk layer) and CachedRetriever (a Retriever wrapper that loads/stores the fitted index keyed by a corpus fingerprint, transparently — warm loads score byte-identically to a cold fit). Pass via Router(retriever=CachedRetriever(TfIdfRetriever(), cache)). Codec + fingerprint live in routing/_index_codec.py. |
routing/_index_codec.py |
Private helper for index_cache.py: index_fingerprint() (deterministic ordered-corpus SHA-256) + the IndexCodec contract and bundled TFIDF_CODEC. Not public API; names re-exported from index_cache. |
routing/trace.py |
RouteTrace + TraceStep structured routing audit (issue #51) |
routing/explanation.py |
RouteResult.explanation() Markdown / dict rendering (issue #226) |
routing/pipeline.py |
RoutingPipeline composer — explicit retrieve → rerank → navigate → pack stages (issue #56) |
routing/navigator.py |
BeamSearchNavigator (lifted from router.py) + rank_collected — the score-sort/active-filter helper is re-exported from routing/__init__.py for custom Navigator implementations (issues #56, #288) |
routing/packer.py |
DefaultCardPacker wrapping make_choice_cards for the pipeline pack stage (issue #56) |
routing/history.py |
RouteHistory dataclass + adjust_scores (history-aware re-routing, issue #27) |
routing/feedback.py |
Optional feedback-aware routing scores (issue #318): ExecutionFeedback (contextweaver-native, not a weaver-spec type), DeterministicScoreProvider (default no-op), FeedbackAwareScoreProvider, aggregate_feedback. Plugs into Router(score_provider=...); default None keeps routing deterministic. |
routing/selection.py |
Structured route→select contract (issues #515/#479), both pure/deterministic: selection_schema emits the routed candidate IDs as a provider-native constrained-selection schema (json_schema/openai/anthropic) so a model can only pick a routed tool_id ("constrain before"); validate_selection → SelectionValidation validates/repairs (strip → case-fold → unique-prefix; ambiguous matches rejected, never guessed) a returned ID against the candidates ("validate after"). Surfaced on RouteResult.selection_schema() / RouteResult.validate_selection(); to_routing_decision resolves + records the outcome. Shortlist composition (pin_ids always-include + per-namespace namespace_quota, issue #509) lives in routing/filters.compose_shortlist and is exposed via Router.route(...). |
extras/embeddings.py |
SentenceTransformerBackend + HybridEmbeddingRetriever + HashingEmbeddingBackend (re-exported) behind the [embeddings] extra (issue #8) |
extras/embeddings_hashing.py |
HashingEmbeddingBackend — stdlib-only deterministic EmbeddingBackend using blake2b hashing trick; no extras required (issue #266) |
_schema_gen.py |
Dataclass → JSON Schema (Draft 2020-12) generator + make schemas-check engine (issue #225) |
routing/tool_id.py |
Canonical tool_id grammar (parse_tool_id / format_tool_id / compute_hash8) per docs/gateway_spec.md §1 |
routing/primitive_id.py |
Unified cross-primitive identity + collision policy for tools/resources/prompts (parse_primitive_id / format_primitive_id / canonical_resource_id / canonical_prompt_id / resolve_collisions) per docs/gateway_spec.md §9. Tools keep the bare tool_id; resources/prompts get disjoint kind:: ids (issue #671). |
routing/path.py |
tool_browse path-navigation grammar (parse_path / resolve_path) per docs/gateway_spec.md §3 |
routing/hydration.py |
Public schema-hydration helpers — SchemaSource (from raw dict / JSON file / MCP tools-list), hydrate_with_schema, lazy_schema_resolver. Reference architectures use these to resolve a tool's full input schema from a sidecar source rather than hand-rolling a _FULL_SCHEMAS dict. Inline args_schema on the catalog item wins; sidecar only fills empties. Issue #261. |
adapters/ |
MCP, FastMCP, A2A, weaver-spec, CrewAI, Pydantic AI, smolagents, Agno, LangChain, OpenAI Agents SDK, Google ADK, Microsoft Agent Framework, OpenAPI, Agent Skills protocol adapters + MCP proxy / gateway runtime + provider-message ingestion helpers for OpenAI / Anthropic / Gemini chat histories. Framework tool-catalog adapters share adapters/_framework_common.py (issue #454). (issues #13, #28, #29, #34, #193, #194, #219, #222, #272, #274, #275, #430, #454, #501, #502, #545, #546, #547) |
adapters/chainweaver.py |
ChainWeaver flow-export → SelectableItem(kind="flow") import (chainweaver_flow_to_selectable, chainweaver_flows_to_catalog, load_chainweaver_export, issue #334). Pure data; no ChainWeaver dependency. Preserves name/description/input+output schemas; stamps metadata["runtime"]="chainweaver" + flow id/version. |
adapters/crewai.py |
CrewAI BaseTool (or equivalent plain-dict shape) ↔ SelectableItem (crewai_tool_to_selectable, crewai_tools_to_catalog, infer_crewai_namespace, load_crewai_catalog, issue #193) |
adapters/pydantic_ai.py |
Pydantic AI Tool ↔ SelectableItem and ModelMessage ↔ ContextItem lossless round-trip (pydantic_ai_tool_to_selectable, pydantic_ai_tools_to_catalog, load_pydantic_ai_catalog, from_/to_pydantic_ai_messages, issue #272) — heavy decode/encode helpers live in adapters/_pydantic_ai_messages.py |
adapters/smolagents.py |
Hugging Face smolagents Tool ↔ SelectableItem and MultiStepAgent.memory.steps → ContextItems (smolagents_tool_to_selectable, smolagents_tools_to_catalog, load_smolagents_catalog, from_smolagents_agent, issue #274) |
adapters/agno.py |
Agno (formerly Phidata) Function / Toolkit ↔ SelectableItem and AgentSession → ContextItems (agno_tool_to_selectable, agno_tools_to_catalog, load_agno_catalog, from_agno_session, issue #275) |
adapters/_framework_common.py |
Shared, framework-agnostic conversion scaffolding for the framework tool-catalog adapters (issue #454): infer_namespace, strip_namespace_prefix, coerce_schema_dict, collect_tags, require_name_description. Pure/stateless, imports no framework lib. Private — not exported. New adapters reuse these instead of re-implementing namespace/schema/tag mechanics. |
adapters/langchain.py |
LangChain BaseTool (or equivalent plain-dict shape) ↔ SelectableItem (langchain_tool_to_selectable, langchain_tools_to_catalog, infer_langchain_namespace, load_langchain_catalog, issue #502). [langchain] extra for live loading; plain-dict path needs no extra. |
adapters/openai_agents.py |
OpenAI Agents SDK FunctionTool ↔ SelectableItem and run items → ContextItems (openai_agents_tool_to_selectable, openai_agents_tools_to_catalog, load_openai_agents_catalog, from_openai_agents_run, issue #501). Run-item ingestion lives in adapters/_openai_agents_run.py. [openai-agents] extra for live loading. |
adapters/google_adk.py |
Google ADK tools ↔ SelectableItem and Session.events → ContextItems (google_adk_tool_to_selectable, google_adk_tools_to_catalog, load_google_adk_catalog, from_google_adk_session, issue #547). Session ingestion lives in adapters/_google_adk_session.py. [google-adk] extra for live loading. |
adapters/agent_framework.py |
Microsoft Agent Framework (AutoGen / Semantic Kernel lineage) tools ↔ SelectableItem and thread ChatMessages → ContextItems (agent_framework_tool_to_selectable, agent_framework_tools_to_catalog, load_agent_framework_catalog, from_agent_framework_thread, issue #430). Thread ingestion lives in adapters/_agent_framework_thread.py. [agent-framework] extra for live loading. |
adapters/openapi.py |
OpenAPI 3.0/3.1 operations → SelectableItem catalog (openapi_operation_to_selectable, openapi_spec_to_catalog, load_openapi_catalog, infer_openapi_namespace, issue #546). Routes over REST APIs; never calls them. Local $ref resolution + parameters/requestBody → args_schema composition + method→safety tags live in adapters/_openapi_schema.py. No extra — PyYAML/jsonschema are core. |
adapters/agent_skills.py |
Agent Skills (SKILL.md) directories → kind="skill" SelectableItems with lazy body hydration (skill_to_selectable, load_skills_catalog, parse_skill_frontmatter, SkillBodySource, issue #545). Frontmatter routes; SkillBodySource resolves the body/resources on selection (mirrors routing/hydration.SchemaSource). No extra — PyYAML is core. |
adapters/_okf_io.py / _okf_materialize.py / _okf_coerce.py |
Private OKF-style Markdown-plus-YAML-frontmatter parsing core shared by the four knowledge-source adapters below (issues #736/#763/#767/#776): permissive frontmatter parsing (never raises — degrades to a LoadDiagnostic), the KnowledgeNode dataclass, deterministic directory walk, and materialisation into ContextItem (kind=doc_snippet, provenance under the _contextweaver metadata namespace) + deterministic relevance scoring. _okf_coerce.py holds the pure frontmatter value-coercion helpers (scalar/float/expiry coercion + JSON-safe normalisation of dates/bytes, so the to_dict/ContextItem.metadata "JSON-compatible" contract holds), split out to isolate frontmatter coercion from parsing and materialisation. Not public API; not re-exported at the adapters package level (mirrors mcp_primitives/gateway_primitives — a related family kept at submodule level to stay within adapters/__init__.py's frozen size ceiling). |
adapters/okf.py |
OKF bundle loader as a context source (issue #736): load_okf_bundle, okf_nodes_to_context_items, select_knowledge. index.md/log.md are bundle metadata/history, excluded from concept content by default. |
adapters/repo_knowledge.py |
Repository-knowledge bundles as context sources (issue #763): load_repo_knowledge (plain-Markdown fallback, max_files/max_total_bytes guardrails, links never auto-followed), classify_usage (deterministic usage tags — plain metadata strings, not Phase values; extending Phase is issue #587's separate concern), select_repo_knowledge. |
adapters/lessons.py |
LessonWeaver lesson bundles as lifecycle-aware context sources (issue #767): load_lesson_bundle, LessonSelectionPolicy (excludes rejected/deprecated/unreviewed-candidate lessons by default), eligible_lessons, select_lessons — every exclusion is reported with a reason, never silent. |
adapters/expertise_pack.py |
ExpertisePacks as bounded context sources (issue #776): load_expertise_pack validates pack structure (an index.md declaring version, every node carrying a key) — the canonical schema is tracked externally at dgenio/weaver-spec#184 and not yet bound. detect_conflicts is deterministic-only (literal same-key text disagreement, scoped by task_tags/expiry) — no LLM-backed contradiction inference. |
adapters/proxy_runtime.py |
ProxyRuntime shared core + ExposureMode enum + UpstreamCall Protocol (issue #29) |
adapters/gateway_diagnostics.py / gateway_catalog_diagnostics.py |
Sanitized ProxyRuntime instrumentation plus exact gateway/proxy static-schema exposure calculations: catalog, browse/hydrate/execute/view events, savings, artifact-view usage, and latency (issues #370/#378). |
adapters/mcp_gateway.py |
Two-tool gateway dispatch (tool_browse + tool_execute + tool_view, issues #28 / #34) |
adapters/mcp_proxy.py |
Transparent proxy dispatch (stripped tools/list + tool_hydrate + tool_execute, issue #13) |
adapters/mcp_upstream.py |
Concrete UpstreamCall adapters (StubUpstream, McpClientUpstream, MultiplexUpstream) |
adapters/_config_coerce.py |
Private config-value coercion helpers (interpolate_env for ${env:VAR}, coerce_bool, opt_positive_number, str_tuple, str_map) shared by upstream_config.py / startup_policy.py / artifact_policy.py (issue #366). Not public API. |
adapters/upstream_config.py |
Pure-data config for one live upstream MCP server (issues #366/#368): UpstreamSpec (type: stdio|http|sse, command/args/env or url/headers, namespace, required, include_tools/exclude_tools globs, timeout), parse_upstreams_config. |
adapters/startup_policy.py |
Fault-tolerant multi-upstream startup policy (issue #374): StartupPolicy (mode: degraded|strict, upstream_timeout_seconds, min_healthy_upstreams, fail_on_empty_catalog), UpstreamStatus, StartupReport, detect_tool_name_collisions (diagnostic only — routing still resolves collisions first-registered-wins via MultiplexUpstream). |
adapters/artifact_policy.py |
Artifact lifecycle policy for mcp serve --state-dir (issue #375): ArtifactPolicy (ttl_seconds, max_bytes, max_artifacts, redact_secrets), wired into JsonFileArtifactStore. |
adapters/upstream_launch.py |
Live multi-upstream launch behaviour (issues #366/#368/#374): launch_upstreams connects every configured UpstreamSpec under AsyncExitStack, classifies failures via gateway_error.classify_upstream_exception, and composes survivors behind MultiplexUpstream; NamespacedFilteredUpstream applies namespace-prefixing + include/exclude filtering at the list_tools boundary. Connect-step timeouts bound only session.initialize() (a self-contained RPC) — never the stdio_client/ClientSession context-manager entry itself, since wrapping that in asyncio.wait_for schedules a new Task and violates anyio's same-task cancel-scope invariant when AsyncExitStack closes it later from the caller's task. |
adapters/mcp_gateway_server.py |
Bind mcp_gateway onto mcp.server.Server over stdio (issue #28); optional primitive_runtime= also advertises/dispatches the four resource/prompt meta-tools (issues #669/#670) |
adapters/mcp_primitives.py |
MCP resource/prompt → SelectableItem(kind="resource"/"prompt") converters + resources/read / prompts/get result→envelope wrappers (issues #669/#670). Emits ids via routing/primitive_id. |
adapters/gateway_primitives.py |
PrimitiveGatewayRuntime + PrimitiveUpstream Protocol — bounded-choice routing + firewall for resources/prompts, sharing the tool runtime's ContextManager (issues #669/#670/#555). |
adapters/_primitive_index.py |
Private single-kind catalog+graph+router+browse helper for gateway_primitives (isolates catalog/graph/router/browse mechanics from adapter wrappers). Not public API. |
adapters/mcp_gateway_primitives.py |
The four resource/prompt gateway meta-tools (resource_browse / resource_read / prompt_browse / prompt_get) + dispatch, mirroring mcp_gateway (issues #669/#670). |
adapters/mcp_primitive_upstream.py |
Concrete PrimitiveUpstream adapters mirroring mcp_upstream: StubPrimitiveUpstream (in-process), McpClientPrimitiveUpstream (wraps an MCP ClientSession), MultiplexPrimitiveUpstream (multi-server fan-out). Transport errors raise (the runtime classifies them) per the Protocol contract (issues #669/#670). |
adapters/mcp_proxy_server.py |
Bind mcp_proxy onto mcp.server.Server over stdio (issue #13) |
adapters/sidecar_contract.py |
HTTP sidecar wire contract (issue #674): RouteRequest/RouteResponse/CompactRequest/CompactResponse/SidecarError dataclasses + SIDECAR_API_VERSION. Pure, dependency-free; the published JSON Schemas live under schemas/sidecar/v1/. |
adapters/sidecar.py |
HTTP sidecar runtime (issue #675/#676): SidecarConfig + SidecarApp.dispatch — transport-free (method, path, headers, body) → (status, json) over the sync Router (/v1/route) and compact_tool_result facade (/v1/compact). Optional bearer-token auth, per-client rate limiting (reuses gateway_controls.RateLimiter), body-size cap, and typed SidecarError responses; never raises across the HTTP boundary. |
adapters/_sidecar_http.py |
Stdlib http.server.ThreadingHTTPServer binding for SidecarApp (issue #675). No third-party dependency. Public re-exports: serve_api (blocking serve) + make_sidecar_server (build-only, for tests). Not public API itself. |
adapters/_sidecar_validation.py |
Stateless parsing + field-validation helpers shared by sidecar.py and sidecar_contract.py (request-body JSON decode, bearer-token extraction, typed contract-field coercions). Pure, dependency-free; raises ConfigError on malformed input. Not public API. |
adapters/gateway_error.py |
Structured GatewayError (codes + §3.4 wire shape) + retryable hint. Upstream-error taxonomy: classify_upstream_exception maps timeouts/connection/auth/permission/rate failures to UPSTREAM_TIMEOUT/UPSTREAM_UNAVAILABLE/AUTH_FAILED/PERMISSION_DENIED/RATE_LIMITED (fallback UPSTREAM_ERROR); redact_upstream_detail strips control chars + caps length on model-visible detail (issue #485). |
adapters/gateway_validation.py |
Untrusted-schema hardening for the gateway ingest path (issues #464/#484): SchemaLimits/SchemaFinding/SkippedTool/CatalogRefreshReport, check_schema_health (meta-validation + iterative size/depth/property bounds), build_validator (cached per tool_id). Pure, deterministic; iterative traversal avoids stack exhaustion on hostile schemas. |
adapters/gateway_args.py |
Opt-in deterministic tool-call argument repair (issue #488): normalize_args (stringified-object parse + schema-demanded str→int/number/boolean/null coercion) + Repair. Gated behind ProxyRuntime(tolerant_args=True); never renames keys, drops keys, or fuzzy-matches. |
adapters/gateway_policy.py |
Pure-data config + result types for the dispatch-path controls (issues #529/#482/#483): RetryPolicy (bounded backoff), RateLimit/RateLimitPolicy (per-session quotas), DryRunReport. All defaults inert; to_dict/from_dict for mcp serve --config. |
adapters/gateway_controls.py |
Runtime mechanisms behind gateway_policy (issues #529/#482/#512): call_with_retry (retry loop, injectable sleep), RateLimiter (sliding-window + cumulative counters, injectable clock), ToolResultCache (TTL+LRU read-only response cache). All opt-in; wired into ProxyRuntime.execute/browse/view. |
adapters/gateway_authz.py |
Runtime authorization / policy gate (issue #373): ToolPolicy (ordered PolicyRules + default), PolicyContext, PolicyDecision, policy_gate_error. Pure/deterministic (first-match, case-sensitive globs). Evaluated before tool_execute dispatch and — via meta_tool="tool_view" rules — before tool_view raw egress (issue #746). Default allow (inert); deny→POLICY_DENIED, require_approval→AUTH_REQUIRED. to_dict/from_dict for the policy block of mcp serve --config. |
adapters/gateway_visibility.py |
Audience-scoped catalog visibility profiles (issue #379): VisibilityProfile (glob include/exclude over namespaces/domains, risk/lifecycle exclusion, side-effect + environment allowlists), evaluate_visibility/filter_catalog (denials carry reasons), parse_profiles for a visibility_profiles: config block. Mirrors gateway_authz's first-match pattern; exclude rules fail open on unknown inventory metadata, allowlists fail closed. Applied at serve startup before ChoiceCards. |
adapters/catalog_pin.py |
Catalog pinning against tool-surface drift (issue #656): PinPolicy (expected_hash, mode: warn|strict), check_catalog_pin via routing.manifest.compute_catalog_hash, enforce_pin (strict-mode mismatch raises with both hashes + re-pin hint). pin: block of mcp serve --config. |
adapters/live_refresh.py |
Notification-driven live catalog refresh (issue #424): LiveRefreshPolicy + LiveRefresher turn upstream tools/list_changed into ProxyRuntime's atomic register_tool_defs_sync — debounced, sliding-minute rate-limited, inert by default; make_message_handler wires it to the ClientSession message handler threaded through launch_upstreams. |
adapters/serve_lifecycle.py |
Graceful shutdown for the serve loops (issue #626): ShutdownController — SIGINT/SIGTERM capture (graceful where loops lack signal support), drain-then-cancel of in-flight work, defensive flush/close of stores and sinks with per-object error collection → ShutdownReport. Transport-agnostic. |
adapters/memory_server.py |
Standalone MCP memory server (issue #632): build_memory_server / run_memory_server_stdio expose episodic/fact stores as four MCP tools (memory_add_episode/search_episodes/put_fact/get_facts) with deterministic content-hash ids, the gateway ARGS_INVALID error shape, and opt-in read-path secret scrubbing via the shared contextweaver.secrets helpers. Backs mcp memory-serve. |
adapters/gateway_scorecard.py |
Tool-surface health scorecard (issue #380): build_scorecard joins inventory metadata (#377) with diagnostic-event streams (#382) into a deterministic report (usage/latency/failure hot-spots, unused/deprecated-in-use, collision counts via #381) + Markdown/CSV/JSON renders. Backs mcp scorecard. |
adapters/gateway_status.py |
Gateway status surface for mcp status (issue #655): GatewayStatus snapshot + rate-limited atomic StatusWriter (injectable clock, coalescing, force()), read_status (ConfigError with --state-dir hint), render_status (uptime + 30s staleness warning). |
adapters/gateway_doctor.py / _doctor_checks.py |
Gateway preflight for mcp doctor (issue #395): run_doctor → DoctorReport of never-raising DoctorFindings — config/key-set/XOR checks, catalog load + references + weak-metadata, ChoiceCard schema-hiding probe, hydration, artifact-store writability, optional-extras info, opt-in live upstream launch + routing smoke queries. Check bodies live in the private _doctor_checks.py (size split). |
adapters/sampling_call_fn.py |
MCP sampling-backed call_fn (issue #623): make_sampling_call_fn / make_async_sampling_call_fn bridge the firewall's Summarizer seam to the connected client's model via sampling/createMessage. Sync variant runs on a worker thread through the server loop with a deadlock guard + timeout; non-text results raise so consumers fall back to the deterministic path. Opt-in only. |
adapters/wire_capture.py |
Record/replay harness pinning gateway protocol behaviour with committed golden transcripts (issue #654): WireRecorder / save_transcript / load_transcript / replay_and_verify at the dispatch boundary (JSON-RPC framing is SDK-owned; content bodies JSON-decoded for field-level diffs, volatile paths normalised). |
adapters/gateway_presets.py |
Named gateway policy presets (issue #664): GatewayPreset.from_preset("safe"|"balanced"|"throughput") bundles a ToolPolicy + RetryPolicy + RateLimitPolicy + CacheConfig (also defined here — pure-data config for gateway_controls.ToolResultCache). Selectable via mcp serve --policy-preset / policy_preset config key; an explicit policy/retry/rate_limits/cache block wins over the preset for that block. Deterministic to_dict() backs mcp serve --print-effective-policy. |
adapters/_proxy_dispatch.py |
Private dispatch helpers keeping proxy_runtime.py within its size ceiling: persist_result_artifacts, execute_policy_error/view_policy_error (policy-context construction + gate call), rate_limited_error, unverified_annotations, build_dry_run_report, UpstreamNameIndex. Not public API. |
adapters/openai_messages.py |
OpenAI Chat Completions messages ↔ ContextItem round-trip (from_/to_openai_messages, issue #219) |
adapters/anthropic_messages.py |
Anthropic Messages API messages ↔ ContextItem round-trip (from_/to_anthropic_messages, issue #222) |
adapters/gemini_contents.py |
Google Gemini contents[] ↔ ContextItem round-trip (from_/to_gemini_contents, issue #222) |
extras/otel.py |
OpenTelemetry GenAI integration (OTelEventHook — invoke_agent / execute_tool spans + GenAI SemConv attributes, gated behind the [otel] extra, issue #224). |
extras/llm_summarizer.py |
Optional LlmSummarizer / LlmExtractor — LLM-backed Summarizer / Extractor plugins for the firewall. Take a user-supplied call_fn (no LLM SDK dep, no extra) and degrade to the rule-based path on any failure (issue #26). Optional provider_metadata surfaces on FirewallStats.llm_provider for audit; truncated inputs carry a deterministic omission marker (issue #384). |
extras/llm_guard.py |
GuardedCallFn — policy envelope for user-supplied LLM call_fn callables (issue #494): GuardPolicy (call cap, consecutive-failure circuit breaker with cooldown/half-open trial, timeout accounting — post-hoc by default, opt-in thread-enforced hard timeout) + live GuardStats. Injectable monotonic clock. Rejections raise PolicyViolationError so plugins degrade to their deterministic fallbacks. No LLM SDK dep. |
extras/catalog_enrich.py |
Offline LLM-assisted catalog metadata enrichment (issue #383): enrich_catalog emits reviewable EnrichmentSuggestions from metadata-only prompts (enum values validated against #377 vocabularies; malformed output + guard rejections collected as skips), never mutating the catalog; apply_suggestions is the explicit reviewed opt-in. |
extras/ranker.py / _ranker_examples.py |
Telemetry-trained tool reranker behind the [ranker] extra (issue #388): deterministic featurize, sklearn-gated ToolRanker (fixed-seed fit, JSON coefficient + model-card persistence — no pickle, pure-Python prediction so saved models load without sklearn), evaluate_ranker vs a lexical baseline. Example derivation from DiagnosticEvents in the private sibling. |
extras/model_setup.py |
Local embedding-model setup helpers behind models doctor / models download (issue #386): EmbeddingModelConfig (the models.embeddings block), network-free run_model_doctor (extra/torch/cache/model-presence checks with install hints), and an explicit download_model (the only network-touching call). |
extras/memory/ |
External-memory backend adapters that implement EpisodicStore / FactStore against an existing long-lived memory deployment without widening the Protocols (issue #195). |
extras/memory/mem0.py |
Mem0EpisodicStore + Mem0FactStore — wrap a mem0.Memory instance scoped by user_id; writes go through Memory.add(infer=False) and items are stamped with cw_episode_id / cw_fact_id metadata for canonical-ID resolution. Gated behind the [mem0] extra (issue #195). |
extras/memory/zep.py |
ZepEpisodicStore + ZepFactStore — wrap a zep_cloud.Zep client scoped by user_id; persist items as JSON graph episodes (graph.add(type="json")) stamped with cw_* IDs, resolving back via graph.episode.get_by_user_id. Episodic search is client-side (Zep graph search is edge/node-shaped). Gated behind the [zep] extra (issue #195). |
extras/memory/_zep_common.py |
Internal helpers backing zep.py (isolates shared Zep scope, scan, coercion, and write mechanics): shared cw_* constants, the ZepBackendError exception, the JSON/scan helpers (_episode_records / _episode_uuid / _episode_payload), the defensive payload-coercion helpers (_coerce_str_tags / _coerce_metadata), and the _ZepStoreBase scope/scan/write base. Carries the same [zep]-extra import guard (issue #195). |
extras/memory/langmem.py |
LangMemEpisodicStore + LangMemFactStore — wrap any LangGraph BaseStore scoped by a namespace tuple; canonical ID is the store key, value is the dataclass to_dict() payload (direct, lossless KV). search delegates to BaseStore.search. Gated behind the [langmem] extra (issue #195). |
eval/ |
Evaluation harness (issue #12): EvalCase / EvalDataset (gold datasets), evaluate_routing → RoutingEvalReport (top-k recall, MRR, confidence gap, beam steps), evaluate_context → ContextEvalReport (budget utilisation + token savings vs naive concat). Pure-stdlib, deterministic; backs the eval CLI subcommand. |
eval/consolidation.py |
Consolidation quality evaluation harness (issue #683): evaluate_consolidation → ConsolidationEvalReport (precision / coverage against an optional gold set + dedup ratio). Pure-stdlib, offline, deterministic. |
eval/metrics.py |
Canonical rank-based routing metrics — recall_at_k (classic fractional recall@k), precision_at_k, reciprocal_rank (issue #354). Single source of truth imported by both eval/routing.py and benchmarks/benchmark.py so the harness and the benchmark script can no longer define the same names with different semantics. |
__main__.py |
CLI: 15 top-level subcommands (start, demo, build, route, print-tree, init, ingest, replay, stats, inspect, budget-check, eval, consolidate, verify, serve-api) plus the mcp and catalog Typer sub-apps. start prints deterministic deployment-intent guidance without side effects (issue #660); inspect renders payload-safe context/routing/artifact reports (issue #398); catalog lint surfaces normalization + reference findings with --json and CI exit codes (issue #538). |
_mcp_cli.py |
Backs the mcp Typer sub-app. Hosts mcp serve, mcp inspect, mcp stats, mcp generate-configs, mcp incident-pack, and mcp import-vscode; accepts native contextweaver, raw MCP tools/list, and {tools:[...]} catalog shapes. mcp serve --diagnostics FILE appends sanitized JSONL and --quiet suppresses lifecycle stderr; both are config-file keys. mcp serve --state-dir DIR (config key state_dir) persists gateway state — events.sqlite3 + artifacts/ — so artifact handles and event history survive a restart (issue #511); omit it for the in-memory default. An artifacts: config block (issue #375) wires TTL/quota/redaction into that store. mcp generate-configs emits deterministic multi-client recipe artifacts from one canonical mcp serve --config input (issue #659). mcp incident-pack creates local redacted triage bundles and never reads shell history automatically (issue #661). mcp serve --policy-preset (config key policy_preset) selects a named GatewayPreset (issue #664); an explicit policy/retry/rate_limits/cache block still wins over the preset for that block. mcp serve --print-effective-policy prints the resolved bundle as JSON and exits without requiring the catalog to exist on disk. The default serve path remains a static catalog + stub upstream; an upstreams: config block (issues #366/#368/#374) switches to _serve_live(), which launches real upstream MCP servers via adapters.upstream_launch.launch_upstreams under a startup: fault-tolerance policy — resources/prompts are not yet supported over live upstreams. mcp import-vscode (issue #367) migrates a VS Code MCP config into an upstreams: gateway config via _vscode_import.py; defaults to --dry-run. |
data/ |
Packaged data files shipped inside the wheel via [tool.setuptools.package-data]. Exposes gateway_catalog_path() (resolves mcp_gateway_catalog.yaml to a concrete Path for both editable installs and zipped wheels — falls back to a persistent cache under tempfile.gettempdir()/contextweaver/ for zipimport). Issue #264. |
examples/recipes/ |
MCP-client integration recipes: installed-CLI configs for Claude Desktop, Claude Code, GitHub Copilot, and Cursor plus gateway_config.yaml; serve_gateway.py remains a legacy/custom-runtime launcher (issues #278, #279, #346, #371, #429, #437). |
Changes
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 413 lines · 17,577 tokens per session scan A efe713695b91
contextweaver AGENTS.md is an instructions file published in the GitHub repository dgenio/contextweaver (9 stars, last pushed 2d ago), licensed Apache-2.0. It adds 17,577 tokens to every session, about $0.0879 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Related
Other instructions, from other repositories
kindex CLAUDE.md
Instructions for jmcentire/kindex, covering kindex — knowledge graph for ai-assisted workflows, running tests, all tests, specific module and with coverage.
ChainWeaver AGENTS.md
Instructions for dgenio/ChainWeaver, covering chainweaver — agent instructions, 1. project identity, 2. domain vocabulary, 3. repository layout and 4. core invariants.
ChainWeaver copilot-instructions.md
Instructions for dgenio/ChainWeaver, covering copilot instructions — chainweaver, scoped guidance, review-critical rules, executor guardrails and vocabulary.
ChainWeaver testing.instructions.md
Instructions for dgenio/ChainWeaver, covering testing instructions — chainweaver, framework, structure, fixtures and coverage patterns.
ChainWeaver python-source.instructions.md
Instructions for dgenio/ChainWeaver, covering python source instructions — chainweaver, module conventions, pydantic patterns, exception patterns and export rules.
ChainWeaver chainweaver.instructions.md
Instructions for dgenio/ChainWeaver: Also read the nearest path-scoped AGENTS.md for the subtree you are changing — the index is in AGENTS.md § 11.